 | What can we learn from errors in human-human and wizard-of-Oz systems that will help us to handle error in human-machine dialogue systems? |
| How do systems detect when a dialogue is `going wrong'? How do they define such conditions? What factors are the key contributors to and indicators of `bad' dialogues? |
| How do systems identify their own errors? What are the most important causes of such errors, from the user side (e.g. non-native accent, hyperarticulated speaking style, gender, age) and from the system side (e.g. inappropriate prompts)? How difficult is it to determine the causes of particular error? |
| How can we predict which dialogues will be successful? How should we define `success'? What features can best predict it? |
| What mechanisms can be devised to allow systems to recover from error gracefully? Can we devise adaptive strategies to identify patterns of error and respond accordingly? |
| What sorts of behavior do users exhibit when faced with system errors? Can these be taken into account in error handling? |
| What measures (better prompts, anticipation of likely error, better help information) can be taken to minimize possible errors? |
